CNN and LSTM Models for fMRI-based Schizophrenia Classification Using c-ICA of dFNC

M. Moein Esfahani, Hemanth Venkateswara, Zening Fu, Ram Ballem, Vince D. CalhounmedRxiv 2025 · DOI

Abstract

Resting-state fMRI (rs-fMRI) captures brain activity at rest, demonstrating how different regions interact without explicit task-based influences. This provides insights into both healthy and disordered brain states. However, clinical application of rs-fMRI remains challenging due to the wide variability in functional connectivity across individuals. Traditional data-driven methods like independent component analysis (ICA) struggle to balance these individual differences with broader patterns. Constrained methods, such as constrained ICA (cICA), have been introduced to address this by integrating templates from multiple external datasets to enhance accuracy and consistency. In our study, we analyzed rs-fMRI data from 100,517 individuals from diverse datasets, processed through a robust quality-control dynamic connectivity pipeline. Using the resulting brain state templates as cICA priors, we examined the effectiveness of cICA for schizophrenia classification using a combined CNN and LSTM architecture. Results showed stable classification accuracy (87.6% to 86.43%) for the CNN model, while the LSTM model performed less optimally, likely due to sequence processing, yet still yielded comparable results. These findings underscore the potential of group-informed methods and prior data templates in constrained dynamic ICA, offering improved reliability and clinical relevance in rs-fMRI analysis.

Introduction

Resting-state functional MRI (rs-fMRI) is a powerful tool for mapping brain activity because it doesn't require patients to perform complex tasks. However, using this data for clinical diagnosis is notoriously difficult due to inter-subject variability. Every brain is wired differently, and traditional data-driven methods like Independent Component Analysis (ICA) often struggle to distinguish between unique individual traits and pathological markers of diseases like schizophrenia.

Researchers have developed a deep learning pipeline that uses brain connectivity "templates" derived from over 100,000 individuals to classify schizophrenia with high accuracy. By using constrained Independent Component Analysis (cICA) to guide neural networks, this study proves that "group-informed" priors can significantly stabilize and improve brain-based diagnostics.

Methods

The proposed pipeline moves away from purely data-driven discovery toward a "constrained" approach to improve generalizability.

  1. Constrained ICA (cICA): Instead of letting the algorithm find brain networks from scratch for every subject, spatial templates derived from a massive study of 100,517 individuals were used as "priors." This forces the analysis to identify known, replicable Intrinsic Connectivity Networks.
  2. Dynamic Functional Network Connectivity (dFNC): The team analyzed how these networks interact over time, creating a sequence of connectivity matrices representing dynamic brain states.
  3. Deep Learning Classification: The resulting dFNC states were fed into two primary architectures for binary classification (Schizophrenia vs. Healthy Control):
  4. CNN (Convolutional Neural Network): Designed to treat connectivity matrices like images to extract spatial features of brain "wiring."
    • LSTM (Long Short-Term Memory): Designed to process the temporal sequences of how brain states change over the course of the scan.
Figure 1: The proposed research pipeline, illustrating the flow from fMRI data preprocessing to constrained ICA and deep learning classification.

Datasets and Evaluation

The primary dataset was the BSNIP-2 (Bipolar and Schizophrenia Network on Intermediate Phenotypes), including 1,206 participants across six sites. The final classification models were tested on a balanced evaluation subset of 258 subjects (132 Schizophrenia, 126 Healthy Controls). The "Prior" dataset, used solely for template generation, comprised rs-fMRI data from 100,517 individuals.

Key Results and Insights

The study revealed that spatial patterns are significantly stronger biomarkers for schizophrenia than temporal sequences in rs-fMRI data.

  • CNN Superiority: The CNN model consistently outperformed the LSTM, achieving a top accuracy of 87.6% with constrained ICA. This suggests that the spatial arrangement of active brain regions is a more reliable indicator of pathology than the specific timing of state transitions.
  • Priors as Noise Filters: Using templates from 100,000 people helped "clean" individual data, magnifying key features for the CNN while suppressing individual anatomical noise.
  • Stability via Constraints: While the accuracy jump from unconstrained to constrained ICA was modest (86.43% to 87.6%), the reliability and biological grounding of the features improved, ensuring consistency across different subjects.
Figure 6: Accuracy results comparing CNN and LSTM models across constrained and standard ICA case studies.

Discussion and Future Work

While the CNN was highly successful, the LSTM model performed sub-optimally, often hovering near chance levels. This indicates that the sequential dependencies in dFNC may be less relevant for this specific classification task, or that different temporal features (like time-frequency moments) are required for effective LSTM training.

Future work will focus on expanding the model to include other conditions like Bipolar Disorder and further refining the deep learning layers to push accuracy toward clinical viability. These findings underscore the potential of group-informed methods in advancing our understanding of brain function and improving diagnostic reliability in psychiatry.

Summary updated

Abstract

Resting-state fMRI (rs-fMRI) captures brain activity at rest, it demonstrates information on how different regions interact without explicity task-based influences. This provides insights into both healthy and disordered brain states. However, clinical application of rs-fMRI remains challenging due to the wide variability in functional connectivity across individuals. Traditional data-driven methods like independent component analysis (ICA) struggle to balance these individual differences with broader patterns. Constrained methods, such as constrained ICA (cICA), have been introduced to address this by integrating templates from multiple external datasets to enhance accuracy and consistency. In our study, we analyzed rs-fMRI data from 100,517 individuals from diverse datasets, processed through a robust quality-control dynamic connectivity pipeline established in previous work. Using the resulting brain state templates as cICA priors, we examined the effectiveness of cICA for schizophrenia classification using a combined CNN and LSTM architecture. Results showed stable classification accuracy (87.6% to 86.43%) for the CNN model, while the LSTM model performed less optimally, likely due to sequence processing, yet still yielded comparable results. These findings underscore the potential of group-informed methods and prior data templates in constrained dynamic ICA, offering improved reliability and clinical relevance in rs-fMRI analysis and advancing our understanding of brain function.
BibTeX
@misc{esfahani2025cnn,
  author = {M. Moein Esfahani and Hemanth Venkateswara and Zening Fu and Ram Ballem and Vince D. Calhoun},
  title = {{CNN and LSTM Models for fMRI-based Schizophrenia Classification Using c-ICA of dFNC}},
  year = {2025},
  doi = {10.1101/2025.02.27.25322899},
  howpublished = {medRxiv},
}
Figure